Classifying hadronic objects in ATLAS with ML/AI algorithms

This paper summarizes recent advancements in using constituent-based machine learning architectures, such as graph neural networks and transformers, to classify hadronic final states in the ATLAS Experiment, detailing their performance on simulated and real data while outlining future directions for data-driven and model-independent tagging strategies.

The Gist

Imagine the Large Hadron Collider (LHC) at CERN as the world's most powerful, high-speed particle blender. When it smashes protons together, it creates a chaotic explosion of debris. Most of this debris doesn't fly off as single particles; instead, they clump together into "jets"—like streams of water shooting out from a broken hose, but made of subatomic particles.

The ATLAS experiment is like a massive, ultra-fast camera trying to take a picture of this explosion. The big challenge? Sorting the trash.

In this explosion, you have two main types of "trash" streams:

1. Quark jets: These are like tight, focused streams of water.
2. Gluon jets: These are like wide, messy sprays that spread out more.

Physicists need to tell the difference because sometimes they are looking for a specific "treasure" hidden in the mess, like a W boson or a top quark (which are heavy, rare particles). If they can't tell the difference between a "normal" spray and a "special" spray, they might miss the discovery of a lifetime.

Here is how the paper explains the new, high-tech way ATLAS is sorting this mess using Artificial Intelligence (AI).

1. The Old Way vs. The New Way

The Old Way (The "Summary" Approach):
Previously, scientists looked at the jets and calculated a few summary statistics, like "How wide is the spray?" or "How many particles are in it?" They fed these numbers into a simple decision tree (like a flowchart) to guess what the jet was.

• Analogy: It's like trying to identify a fruit by only looking at its weight and color. You might guess it's an apple, but you could be wrong.

The New Way (The "Constituent" Approach):
The paper describes a shift to Machine Learning that looks at every single particle inside the jet, not just the summary numbers. It treats the jet like a cloud of points (a "point cloud") and uses advanced math to see how those points relate to each other.

• Analogy: Instead of just weighing the fruit, the AI takes a 3D scan of every seed, skin texture, and stem, then compares it to millions of other fruits it has seen before.

2. The New AI Tools (The "Detectives")

The paper introduces four main types of AI "detectives" that are better at solving these cases:

• FC DNNs (The List-Makers): These look at the particles in a fixed order. It's like reading a grocery list. It works, but it's a bit rigid.
• EFN & PFN (The "Shape" Shapers): These are smart because they know that the order of particles doesn't matter, only their positions and energies.
  • Analogy: Imagine a bag of marbles. If you shake the bag, the marbles move around. A "Shape Shaper" AI knows the bag still contains the same marbles even if they are in a different order. It focuses purely on the shape of the cloud.
• GNNs (The Social Networkers): These treat every particle as a "node" in a social network. They ask, "Who is friends with whom?"
  • Analogy: In a crowded room, a Gluon jet is like a chaotic party where everyone is talking to everyone. A Quark jet is like a small, tight-knit group of friends huddled together. The GNN maps these relationships to figure out which "party" it is.
• Transformers (The Attention Experts): Borrowed from how computers understand human language (like Google Translate), these use an "attention mechanism."
  • Analogy: When reading a sentence, you don't pay equal attention to every word; you focus on the important ones. Similarly, the Transformer looks at the jet and says, "Hey, this specific particle over here is the most important clue for identifying this jet," and focuses its energy there.

3. The Results: Better Sorting

The paper shows that these new AI detectives are winning:

• Quark vs. Gluon: The new DeParT algorithm (a Transformer-based detective) is much better at rejecting "gluon noise" than the old methods. It's like having a metal detector that only beeps for gold and ignores the rocks.
• Heavy Objects (W Bosons & Top Quarks): When these heavy particles decay, they create "large-radius" jets (big, messy clouds). The ParT and LundNet algorithms are now the gold standard for spotting these.
  • The "LundNet" Trick: One of the new algorithms (LundNetANN) has a special superpower: it learns to ignore the "weight" of the jet. Why? Because sometimes the computer simulation gets the weight wrong. By ignoring the weight, the AI becomes more honest and reliable, reducing errors in the final results.

4. The Future: Data-Driven Truth

The paper concludes with a look ahead. While these AI models are incredibly powerful, they sometimes rely too much on computer simulations (which are like video games). If the "game physics" are slightly off, the AI might get confused.

The future goal is Data-Driven Optimization.

• Analogy: Instead of training the AI only on a video game simulation, scientists want to train it on the real data coming from the collider. They want the AI to learn from the actual universe, not just a computer model of it.

Summary

In short, ATLAS is upgrading its "trash sorting" system. They are moving from simple flowcharts to sophisticated AI that looks at the entire 3D structure of particle jets. By using "Social Network" maps (GNNs) and "Attention" mechanisms (Transformers), they can now spot rare, heavy particles with unprecedented accuracy, bringing us closer to understanding the fundamental laws of the universe.

Technical Summary

1. Problem Statement

In proton-proton collisions at the Large Hadron Collider (LHC), hadronic jets are the most abundant final-state objects. Accurately classifying these jets is critical for:

• Precision Measurements: Distinguishing between quark-initiated and gluon-initiated jets.
• New Physics Searches: Identifying heavy-flavour jets (containing $b$ or $c$ quarks) and boosted heavy objects like $W$ bosons and top quarks that decay hadronically.
• QCD Studies: Understanding non-perturbative effects and parton shower modeling.

Traditional classification methods relied on high-level substructure observables (e.g., track counts, jet width) combined with Boosted Decision Trees (BDTs). However, these approaches often discard granular detector information and fail to capture complex correlations between jet constituents. The challenge lies in developing algorithms that can utilize the full detector granularity while remaining robust against systematic uncertainties and Monte Carlo (MC) generator dependencies.

2. Methodology

The paper outlines a paradigm shift from observable-based taggers to constituent-based architectures. These models directly ingest the four-momenta (four-vectors) of jet constituents (particles or tracks) rather than pre-calculated summary variables. The key architectural approaches discussed include:

• Fully-Connected Deep Neural Networks (FC DNNs): Process ordered constituent features but lack permutation invariance.
• Energy Flow Networks (EFN) & Particle Flow Networks (PFN): Point-cloud models based on the DeepSets approach. They enforce permutation invariance (the output does not change if the order of input particles changes). EFNs specifically use only infrared and collinear-safe variables.
• Graph Neural Networks (GNNs): Represent jet constituents as nodes in a graph with learned geometric edges. ParticleNet is the primary example cited.
• Transformers: Adapted from Natural Language Processing (NLP), these use attention mechanisms to dynamically weigh the relationships between constituents, allowing the model to focus on the most relevant parts of the jet.

3. Key Contributions and Specific Algorithms

A. Quark/Gluon (q/g) and Flavour Tagging

• DeParT (Dynamically Enhanced Particle Transformer): Currently the state-of-the-art for small-radius ($R=0.4$) jets.
  • Mechanism: Operates on Particle Flow Objects (PFOs), utilizing both kinematic features and relational features via attention blocks to replace traditional BDTs.
  • Innovation: It dynamically enhances the representation of constituents, outperforming FC DNNs, PFNs, and EFNs.
• GN2: A transformer-based algorithm for jet flavour tagging.
  • Mechanism: Uses track-level information and auxiliary tasks (vertex and track-origin classification) to stabilize training.
  • Impact: Demonstrates a factor of ~3 improvement in light-jet rejection compared to the Run-2 baseline taggers (DL1r and DL1d).

B. Boosted Object Tagging ($W$ bosons and Top quarks)

• ParT (Particle Transformer): The state-of-the-art for $W$-jet tagging on large-radius ($R=1.0$) jets.
  • Performance: Trained on $W' \to WZ$ and multi-jet events, it outperforms EFN, PFN, and ParticleNet across various transverse momentum ($p_T$) ranges.
• LundNet & LundNetANN: GNNs that exploit the jet clustering history in the Lund Jet Plane (LJP).
  • Innovation: LundNetANN incorporates adversarial training to decorrelate the tagger output from jet mass. This is crucial for reducing systematic uncertainties in background modeling, as it prevents the tagger from learning specific mass distributions from the MC generator.
• ParticleNet for Top Tagging: Achieves state-of-the-art accuracy for top-quark tagging (identifying the three-prong substructure) across a wide $p_T$ range, though with higher computational cost.

4. Results

The paper presents performance metrics based on ATLAS simulations (13 TeV, Pythia8):

• Quark/Gluon Discrimination: DeParT achieves significantly higher gluon-jet rejection ($\epsilon_g^{-1}$) across a broad $p_T$ range compared to FC networks, PFNs, and EFNs, particularly at high quark identification efficiencies ($\epsilon_q$).
• Flavour Tagging: GN2 shows a massive leap in performance over previous generations, reducing light-jet background by a factor of three while maintaining signal efficiency.
• $W$-Boson Tagging: ParT demonstrates superior background rejection compared to ParticleNet, PFN, and EFN.
• Systematic Robustness:
  • While ML models improve accuracy, they exhibit generator dependence. Performance reductions of up to 40% were observed when varying parton-shower models.
  • LundNetANN successfully decorrelates performance from jet mass, providing a more reliable description of QCD background jets compared to standard LundNet, as shown in mass distribution plots.

5. Significance and Future Outlook

• Unprecedented Accuracy: Constituent-based methods (GNNs and Transformers) have established themselves as the leading architectures for jet tagging, surpassing traditional observables by learning complex, non-linear correlations directly from data.
• Challenges: The primary remaining challenge is systematic robustness. The heavy reliance on MC generators for training introduces model dependence, which can bias physics results if the simulation does not perfectly match reality.
• Future Directions:
  • Data-Driven Optimization: Moving towards strategies that optimize models using real data rather than relying solely on simulation.
  • Model-Independent Strategies: Developing techniques to mitigate generator dependence.
  • Hybrid Models: Combining complementary representations, such as merging constituent-based transformers with clustering-tree-based GNNs (like LundNet), to leverage the strengths of both approaches.

In conclusion, the ATLAS Collaboration has successfully transitioned to advanced AI-driven tagging strategies that significantly enhance the physics potential of the LHC, provided that future work addresses the critical issues of systematic uncertainties and model dependence.